ASAT Project

1
ASAT Project

• Two main research thrusts
• Feature extraction
• Evidence combiner
• Feature extraction
• The classical distinctive features are well
  explored, but not solved.
• Many other waveform features and events can be
  extracted reflecting time properties, spectral
  properties, various vocal tract model parameters,
  glottal features, prosodic events and
  combinations thereof.
• Features may at first glace have little relevance
  to articulatory gestures (modulation products,
  etc.)
• Successful feature sets can then be subject to
  perceptual interpretation.
• This approach was successfully implemented in a
  thesis by Necioglu for speaker characterization

2
ASAT Project

• Feature extraction (contd)
• Statistical characterizations that extract
  recurrent patterns can be the basis for such
  features
• One example useful for ultra-low-bit-rate coding
  Ergodic HMMs that are not phonetically based but
  are useful for pattern extraction.
• Take advantage of segmentation event detectors
  used in the latest speech coders (despite dogma,
  the problem and ASR and speech coding cannot be
  completely orthogonal!)
• Robust feature extraction should have confidence
  measures included
• First steps build a toolbox of feature
  extraction modules.

3
ASAT Project

• Evidence Combining / Fusion
• Events will never be perfectly detected.
• Phonetic/sub-word features are never going to be
  perfectly extracted.
• Features can be fuzzy (e.g., nasalization has
  degrees)
• Reliability is affected by speaking style, the
  channel, the length of the event.
• Error bars can be extremely wide
• Common framework seek to represent confidence
  measures as probabilities for straightforward
  combinations. Do not apply thresholding.
• This will require each detected event and each
  high order feature detected to have individual
  non-linear normalizations trained to before
  overall combination.

4
ASAT Project

• Evidence Combining / Fusion (contd)
• This will require each detected event and each
  high order feature to have individual non-linear
  normalizations trained before overall
  combination.
• Some level of brute force will be required to
  estimate these normalizations for new
  contributors.
• Will begin with simple detectors to verify
  approach
• Will study alternate approaches as reported.